JOURNAL OF CHEMICAL PHYSICS

VOLUME 113, NUMBER 22

8 DECEMBER 2000

A climbing image nudged elastic band method for finding saddle points
and minimum energy paths
Graeme Henkelman
Department of Chemistry 351700, University of Washington, Seattle, Washington 98195-1700

Blas P. Uberuaga
Department of Chemistry 351700, University of Washington, Seattle, Washington 98195-1700
and Department of Physics 351560, University of Washington, Seattle, Washington 98195-1560

Hannes Jonsson
Department of Chemistry 351700, University of Washington, Seattle, Washington 98195-1700

Received 23 August 2000; accepted for publication 10 October 2000

A modification of the nudged elastic band method for finding minimum energy paths is presented.
One of the images is made to climb up along the elastic band to converge rigorously on the highest
saddle point. Also, variable spring constants are used to increase the density of images near the top
of the energy barrier to get an improved estimate of the reaction coordinate near the saddle point.
Applications to CH4 dissociative adsorption on Ir111 and H2 on Si100 using plane wave based
density functional theory are presented. 2000 American Institute of Physics.
S0021-96060071246-3

I. INTRODUCTION

points can be obtained from the energy and frequency of

normal modes at the saddle point and the initial state,6,7

An important problem in theoretical chemistry and condensed matter physics is the calculation of transition rates,
for example rates of chemical reactions or diffusion events.
Most often, it is sufficient to treat the motion of the atoms
using classical mechanics, but the transitions of interest are
typically many orders of magnitude slower than vibrations of
the atoms, so a direct simulation of the classical dynamics is
not feasible. For a process with a typical, low activation energy of 0.5 eV, the computer time required to simulate a
classical trajectory long enough that a single transition event
can be expected to occur is on the order of 104 years on
present day computers. This rare event problem is devastating for direct dynamical simulations, but makes it possible
to obtain accurate estimates of transition rates using a purely
statistical approach, namely, transition state theory TST.14
Apart from the BornOppenheimer approximation, TST relies on two basic assumptions: a the rate is slow enough
that a Boltzmann distribution is established and maintained
in the reactant state and b a dividing surface of dimensionality D-1, where D is the number degrees of freedom in the
system, can be identified such that a reacting trajectory going
from the initial state to the final state only crosses the dividing surface once. The dividing surface must, therefore, represent a bottleneck for the transition.
Since atoms in crystals are usually tightly packed and
the typical temperature of interest is low compared with the
melting temperature, the harmonic approximation to TST
hTST can typically be used in studies of diffusion and reactions in crystals or at crystal surfaces.5 This greatly simplifies the problem of estimating the rates. The search for the
optimal transition state then becomes a search for the lowest
few saddle points at the edge of the potential energy basin
corresponding to the initial state. The rate constant for transition through the region around each one of the saddle
0021-9606/2000/113(22)/9901/4/$17.00

k hTST

init
3N
i i

3N1
i
i

e (E

E init)/k T
B

Here, E is the energy of the saddle point, E init is the local

potential energy minimum corresponding to the initial state,
and the i are the corresponding normal mode frequencies.
The symbol refers to the saddle point. All the quantities
can be evaluated from the potential energy surface, at zero
temperature, but entropic effects are included through the
harmonic approximation. The most challenging part in this
calculation is the search for the relevant saddle point.
A path connecting the initial and final states that typically has the greatest statistical weight is the minimum energy path MEP. At any point along the path, the force
acting on the atoms is only pointing along the path. The
energy is stationary for any perpendicular degree of freedom.
The maxima on the MEP are saddle points on the potential
energy surface. The relative distance along the MEP is a
natural choice for a reaction coordinate, and at the saddle
point the direction of the reaction coordinate is given by the
normal mode eigenvector corresponding to negative curvature.
The MEP often has one or more minima in addition to
the minima at the initial and final states. These correspond to
stable intermediate configurations. The MEP will then have
two or more maxima, each one corresponding to a saddle
point. Assuming a Boltzmann population is reached for the
intermediate metastable configurations, the overall rate is
determined by the highest saddle point. It is, therefore, not
sufficient to find a saddle point. One needs to have a good
enough estimate of the shape of the MEP to be able to assign
the highest saddle point as in Eq. 1 in order to get an
accurate estimate of the rate.
9901

2000 American Institute of Physics

9902

Henkelman, Uberuaga, and Jonsson

J. Chem. Phys., Vol. 113, No. 22, 8 December 2000

Many different methods have been presented for finding

MEPs and saddle points.810 Since a first order saddle point
is a maximum in one direction and a minimum in all other
directions, methods for finding saddle points invariably involve some kind of maximization of one degree of freedom
and minimization in other degrees of freedom. The critical
issue is to find a good and inexpensive way to decide which
degree of freedom should be maximized.
The nudged elastic band NEB method is an efficient
method for finding the MEP between a given initial and final
state of a transition.9,11,12 It has become widely used for estimating transition rates within the hTST approximation. The
method has been used both in conjunction with electronic
structure calculations, in particular plane wave based DFT
calculations see, for example, Refs. 1317, and in combination with empirical potentials.1821 Studies of very large
systems, including over a million atoms in the calculation,
have been conducted.22 The MEP is found by constructing a
set of images replicas of the system, typically on the order
of 420, between the initial and final state. A spring interaction between adjacent images is added to ensure continuity
of the path, thus mimicking an elastic band. An optimization
of the band, involving the minimization of the force acting
on the images, brings the band to the MEP.
An essential feature of the NEB method, which distinguishes it from other elastic band methods,2325 is a force
projection which ensures that the spring forces do not interfere with the convergence of the elastic band to the MEP, as
well as ensuring that the true force does not affect the distribution of images along the MEP. It is necessary to estimate
the tangent to the path at each image and every iteration
during the minimization, in order to decompose the true
force and the spring force into components parallel and perpendicular to the path. Only the perpendicular component of
the true force is included, and only the parallel component of
the spring force. This force projection is referred to as
nudging. The spring forces then only control the spacing
of the images along the band. When this projection scheme is
not used, the spring forces tend to prevent the band from
following a curved MEP because of corner-cutting, and
the true force along the path causes the images to slide away
from the high energy regions towards the minima, thereby
reducing the density of images where they are most needed
the sliding-down problem. In the NEB method, there is
no such competition between the true forces and the spring
forces; the strength of the spring forces can be varied by
several orders of magnitude without effecting the equilibrium position of the band.
Recently, an improved way of estimating the tangent to
the elastic band at each image has been presented.26 This
eliminates a problem which occurred in systems where the
force parallel to the MEP was very large compared with the
restoring force perpendicular to the MEP.9 In such situations
kinks could form on the elastic band and prevent rigorous
convergence to the MEP. We use this new way of estimating
the tangent in the calculations presented here.
While the NEB method gives a discrete representation of
the MEP, the energy of saddle points needs to be obtained by
interpolation. When the energy barrier is narrow compared

with the length of the MEP, few images land in the neighborhood of the saddle point and the interpolation can be
inaccurate.
This communication describes a modification of the
NEB method which gives a precise estimate of the saddle
point at no extra cost as compared with the regular NEB.
II. DFT CALCULATIONS OF DISSOCIATIVE
ADSORPTION

The method presented here has been applied to calculations of CH4 dissociative adsorption on the Ir111 surface
and H2 on the Si100 surface using plane wave based density functional theory DFT.27 The PW91 functional28,29 was
used in combination with ultrasoft pseudopotentials.30 The
energy cutoff was 350 eV in the CH4 /Ir111 calculation and
200 eV in the H2 /Si100 calculation. The calculations were
carried out with the VASP code31 which we have extended to
implement the new method presented here. The calculations
were carried out in parallel on a cluster of workstations. The
NEB method lends itself so well to parallel processing that it
is sufficient to use a regular ethernet connection to transfer
data between the nodes.
III. REGULAR NEB METHOD

An elastic band with N1 images can be denoted by

R0 , R1 , R2 , . . . ,RN , where the end points, R0 and RN , are
fixed and given by the energy minima corresponding to the
initial and final states. The N1 intermediate images are
adjusted by the optimization algorithm.
In the NEB method,9,26 the total force acting on an image is the sum of the spring force along the local tangent and
the true force perpendicular to the local tangent
Fi Fsi E Ri ,

where the true force is given by

E Ri E Ri E Ri i .

Here, E is the energy of the system, a function of all the

atomic coordinates, and i is the normalized local tangent at
image i. The spring force is
Fsi k Ri1 Ri Ri Ri1 i ,

where k is the spring constant. An optimization algorithm is

then used to move the images according to the force in Eq.
2. We have used a projected velocity Verlet algorithm.9
The images converge on the MEP with equal spacing if the
spring constant is the same for all the springs. Typically none
of the images lands at or even near the saddle point and the
saddle point energy needs to be estimated by interpolation.
An example of a NEB calculation is shown in Fig. 1.
The MEP for dissociative adsorption of CH4 on an Ir111
surface has a narrow barrier compared with the length of the
MEP. The molecule is 4 above the surface when the reaction coordinate is 1.0. At the other end, at 0.0, the molecule
has broken up into a H and a CH3 fragment sitting on adjacent on-top sites on the Ir111 surface. The resolution of the
MEP near the saddle point is poor and the estimate of the
activation energy obtained from the interpolation is subject

J. Chem. Phys., Vol. 113, No. 22, 8 December 2000

Finding saddle points and minimum energy paths

FIG. 1. Density functional theory calculations of the minimum energy path

for CH4 dissociative adsorption on a Ir111 surface. The dissociated H and
CH3 fragments sitting on adjacent on-top sites correspond to reaction coordinate of 0.0. The CH4 molecule 4 away from the surface corresponds to
1.0. A regular NEB calculation and a climbing image NEB calculation are
compared, both involving 8 movable images. The regular NEB results in a
low resolution of the barrier, and the interpolation gives an underestimate of
the activation energy. The climbing image NEB brings one of the images
right to the saddle point and gives the activation energy precisely with
insignificant additional computational effort.

to large uncertainty. A force and energy based cubic polynomial interpolation was used between each pair of adjacent
images. This is an example of a system where an intermediate minimum is located along the MEP. In fact, it turns out
that this minimum is deeper than the chemisorbed state at the
0.0 end point. The configuration corresponding to the intermediate minimum has the adsorbed H atom at a bridge site.
IV. CLIMBING IMAGE NEB METHOD

The climbing image NEB CI-NEB method constitutes

a small modification to the NEB method. Information about
the shape of the MEP is retained, but a rigorous convergence
to a saddle point is also obtained. This additional feature
does not add any significant computational effort. After a
few iterations with the regular NEB, the image with the highest energy i max is identified. The force on this one image is
not given by Eq. 2 but rather by
Fi maxE Ri max 2E Ri max
E Ri max 2E Ri max i max i max.

This is the full force due to the potential with the component
along the elastic band inverted. The maximum energy image
is not affected by the spring forces at all.
Qualitatively, the climbing image moves up the potential
energy surface along the elastic band and down the potential
surface perpendicular to the band. The other images in the
band serve the purpose of defining the one degree of freedom
for which a maximization of the energy is carried out. Since
the images in the band eventually converge to the MEP, they
give a good approximation to the reaction coordinate around
the saddle point. As long as the CI-NEB method converges,
the climbing image will converge to the saddle point. Since
all the images are being relaxed simultaneously, there is no
additional cost of turning one of the images into a climbing
image.

9903

The results of a CI-NEB calculation of the CH4 dissociation on Ir111 is shown in Fig. 1. A significantly higher
estimate of the activation energy is obtained than with the
regular NEB, using the same number of images. The computational effort is the same to within 10% CI-NEB not
necessarily being slower. Alternatively, one could have run
a second elastic band between the two images adjacent to the
barrier to get a better estimate of the saddle point energy
from the regular NEB, but this would have required more
force evaluations and, therefore, more computational effort.
The activation energy predicted by the DFT/PW91 calculations is approximately 0.4 eV. This calculation still needs to
be corrected33 for quantum zero point energy, dispersion, and
system size effects before it can be compared to the experimental value32 of 0.28 eV. The MEP is nontrivial because it
involves a large relaxation of the substrate. The Ir atom closest to the CH4 molecule in the transition state is pulled out
from the surface plane by 0.5 . This means that the saddle
point does not lie close to the straight line interpolation between the two end points. A more detailed presentation of
the DFT calculations and comparison with experimental results will be given elsewhere.33
The climbing image is not affected by the spring forces.
Therefore, the spacing of the images will be different on
each side of the climbing image. As it moves up to the saddle
point, images on one side will get compressed, and on the
other side spread out. Two or more climbing images can be
specified if the MEP appears to have two or more high
maxima that are close in energy. The only issue is to have
enough images close to the climbing images to get a good
estimate of the reaction coordinate, since this determines the
climbing direction.
V. VARIABLE SPRING CONSTANTS

Since the saddle point is the most important point along

the MEP, one would typically prefer to have more resolution
in the MEP close to the saddle point than near the end points.
The important issue is to get a good enough estimate of the
tangent to the path near the saddle point, especially when a
climbing image is included. As the images are brought closer
to the saddle point, the approximation of the tangent will
become more accurate. Dissociative adsorption of a molecule on a surface is an example of a process where the MEP
is often highly asymmetric and the barrier region is only a
small fraction of the MEP see Figs. 1 and 2. In such cases,
it is more efficient to distribute the images unevenly along
the MEP.
This can be accomplished by using stronger springs near
the saddle point. Because of the nudging, there is no interference between the spring forces that distribute the images
along the MEP and the true force that brings the elastic band
to the MEP. One is, therefore, free to choose different spring
constants between different pairs of images without affecting
the convergence of the band to the MEP, as long as the
number of images is high enough. We have used a scheme
where the spring constant depends linearly on the energy of
the images, in such a way that images with low energy get
connected by a weaker spring constant

9904

Henkelman, Uberuaga, and Jonsson

J. Chem. Phys., Vol. 113, No. 22, 8 December 2000

regular NEB method first 13 iterations with a small time

step until the magnitude of the force had dropped below 1
eV/ and then 166 iterations with a larger time step in the
projected velocity Verlet algorithm9. The number of force
evaluations needed for the CI-NEB calculation with equal
spring constants was 190, and the number of force evaluations needed for the CI-NEB calculation with variable spring
constants was 178. The difference between these numbers is
not significant but simply reflects slight variations in the way
the system moves on the energy surface towards the MEP.
ACKNOWLEDGMENTS

FIG. 2. Density functional theory calculations of the minimum energy path

for H2 dissociative adsorption on a Si100 surface. The H adatoms sitting
on adjacent Si atoms in a surface dimer correspond to reaction coordinate of
0.0. The H2 molecule 3.8 away from the surface corresponds to 1.0. A
regular climbing image NEB calculation with equal spring constants curve
labeled Fixed Springs is compared with a calculation where the spring
constants are scaled with the energy curve labeled Variable Springs,
arbitrarily shifted by 1.0 eV. Both calculations involve 8 movable images.
The variable spring calculation results in a higher resolution of the barrier
with insignificant additional computational effort.

k i

k maxk

E maxE i
E maxE ref

if E i E ref

k maxk if E i E ref .

Here, E i maxEi ,Ei1 is the higher energy of the two images connected by spring i, E max is the maximum value of E i
for the whole elastic band, and E ref is a reference value for
the energy, defining a minimum value of the spring constant.
We have chosen E ref to be the energy of the higher energy
endpoint of the MEP. This choice ensures that the density of
images is roughly equal near the two end points, even for
highly asymmetric MEPs. The spring constant is, therefore,
linearly scaled from a maximum value k max for highest energy images to a minimum value k maxk for images with
energy of E ref or lower.
By choosing E i to be the higher energy of the two images connected by the spring, the two images adjacent to the
climbing image will tend to be symmetrically arranged
around the saddle point. This is only approximately true because of the compression/stretching of the band on each side
of the climbing image.
Figure 2 shows results of a calculation of the dissociation of a H2 molecule on the Si100 surface. This is an
interesting system because of a long standing discrepancy34
between experimental and theoretical measurements of adsorption and dissociation barriers. A CI-NEB calculation
with equal spring constants is compared with a CI-NEB calculation with variable spring constants. The part of the MEP
where the H2 molecule approaches the Si100 surface is flat
and rather uninteresting. The energy scaling of the spring
constants results in images being pulled up towards the barrier region, thus increasing the resolution of the MEP near
the saddle point at the expense of the less important regions.
The number of force evaluations required to reach convergence to within a tolerance of 0.03 eV/ was 179 for the

This work was funded by the National Science Foundation Grant No. CHE-9710995 and by the Petroleum Research
Fund Grant No. PRF#32626-AC5/REF#104788.
H. Eyring, J. Chem. Phys. 3, 107 1935.
E. Wigner, Trans. Faraday Soc. 34, 29 1938.
3
J. C. Keck, Adv. Chem. 13, 85 1967.
4
P. Pechukas, in Dynamics of Molecular Collisions, edited by W. H. Miller
Plenum, New York, 1976, Part B.
5
A. F. Voter and J. D. Doll, J. Chem. Phys. 80, 5832 1984; 82, 80 1985.
6
C. Wert and C. Zener, Phys. Rev. 76, 1169 1949.
7
G. H. Vineyard, J. Phys. Chem. Solids 3, 121 1957.
8
M. L. McKee and M. Page, Reviews in Computational Chemistry, edited
by K. B. Lipkowitz and D. B. Boyd VCH, New York, 1993, Vol. IV.
9
H. Jonsson, G. Mills, and K. W. Jacobsen, Nudged elastic band method
for finding minimum energy paths of transitions, in Classical and Quantum Dynamics in Condensed Phase Simulations, edited by B. J. Berne, G.
Ciccotti, and D. F. Coker World Scientific, Singapore, 1998, p. 385.
10
G. Henkelman, G. Johannesson, and H. Jonsson, Methods for finding
saddle points and minimum energy paths, in Progress on Theoretical
Chemistry and Physics, edited by S. D. Schwartz Kluwer Academic, New
York, 2000 in press.
11
G. Mills and H. Jonsson, Phys. Rev. Lett. 72, 1124 1994.
12
G. Mills, H. Jonsson, and G. K. Schenter, Surf. Sci. 324, 305 1995.
13
B. P. Uberuaga, M. Levskovar, A. P. Smith et al., Phys. Rev. Lett. 84,
2441 2000.
14
J. Song, L. R. Corrales, G. Kresse, and H. Jonsson, Phys. Rev. B submitted.
15
W. Windl, M. M. Bunea, R. Stumpf et al., Phys. Rev. Lett. 83, 4345
1999.
16
R. Stumpf, C. L. Liu, and C. Tracy, Phys. Rev. B 59, 16047 1999.
17
T. C. Shen, J. A. Steckel, and K. D. Jordan, Surf. Sci. 446, 211 2000.
18
M. Villarba and H. Jonsson, Surf. Sci. 317, 15 1994.
19
M. Villarba and H. Jonsson, Surf. Sci. 324, 35 1995.
20
E. Batista and H. Jonsson, Computational Materials Science in press.
21
M. R. So rensen, K. W. Jacobsen, and H. Jonsson, Phys. Rev. Lett. 77,
5067 1996.
22
T. Rasmussen, K. W. Jacobsen, T. Leffers et al., Phys. Rev. Lett. 79, 3676
1997.
23
R. Elber and M. Karplus, Chem. Phys. Lett. 139, 375 1987.
24
R. Czerminski and R. Elber, Int. J. Quantum Chem. 24, 167 1990; J.
Chem. Phys. 92, 5580 1990.
25
R. E. Gillilan and K. R. Wilson, J. Chem. Phys. 97, 1757 1992.
26
G. Henkelman and H. Jonsson, J. Chem. Phys. 113, 9978 2000, this
issue.
27
P. Hohenberg and W. Kohn, Phys. Rev. 136, B864 1964; W. Kohn and
L. J. Sham, Phys. Rev. 140, A1133 1965.
28
W. Kohn, A. D. Becke, and R. G. Parr, J. Phys. Chem. 100, 12974 1996.
29
J. P. Perdew, in Electronic Structure of Solids, edited by P. Ziesche and H.
Eschrig Akademie, Berlin, 1991.
30
D. Vanderbilt, Phys. Rev. B 41, 7892 1990.
31
G. Kresse and J. Hafner, Phys. Rev. B 47, 558 1993; 49, 14251 1994;
G. Kresse and J. Furthmuller, Comput. Mater. Sci. 6, 16 1996; Phys.
Rev. B 54, 11169 1996.
32
D. C. Seets, C. T. Reeves, B. A. Ferguson et al., J. Chem. Phys. 107,
10229 1997.
33
G. Henkelman and H. Jonsson in preparation.
34
F. M. Zimmermann and X. Pan, Phys. Rev. Lett. 85, 618 2000.
1
2